bluff

name	bluff
description	Verify that the URLs and citations in an answer actually resolve, catching hallucinated or dead links before you trust or ship them. Extract every link (bare and markdown), check each over the network, and flag the ones that don't exist. Use when an LLM produced links/citations and you need to confirm they're real, not plausible-looking fiction.

bluff

call its bluff.

LLMs are fluent liars about URLs. They produce a citation that looks exactly right — correct domain, plausible path, confident tone — and it 404s, or never existed. When this skill is active, don't take a link on faith: pull every URL out of the answer and confirm it resolves before presenting it as a source.

What it checks

• Bare URLs — https://… anywhere in the text.
• Markdown links — the target inside [text](url).
• For each: does it actually resolve? 2xx/3xx = real, anything else (404, DNS failure, timeout, connection refused) = a bluff to flag.

It de-dupes and ignores trailing sentence punctuation, so a URL ending a sentence isn't mangled.

Example

Answer under review: "Per the official spec and https://docs.example.com/v2, the limit is 100."

bluff: ✓ [200] https://example.com/spec — ✗ [404] https://docs.example.com/v2. The second citation does not exist; do not ship it as a source.

Output

Deliver the link check, not a description of running it. Don't narrate running bluff.py — the tool card already shows it. List each URL with its ✓/✗ and status, surface the bluffs first, and stop. No preamble, no closing paragraph.

Companion tool

bluff.py in this folder does this mechanically. Pipe an answer in and it reports each link with a ✓/✗, status code, and error.

• It needs network to actually verify links.
• --dry-run extracts and lists the URLs with no network at all — handy when you just want to see what was cited. Extraction is always offline.
• --patterns FILE (repeatable; or the BLUFF_PATTERNS env var) merges custom extraction patterns. The JSON file may carry url_patterns (extra regexes whose match — group 1 if captured, else the whole match — is checked like a URL) and citation_patterns (regexes for non-URL references like DOIs, listed but not network-checked; reported [cited] / ✓). Built-ins stay the base.